Category
Applied AI, LLM workflows, agents, and tooling.
← All articlesThe naive version worked perfectly on fifty pages and failed four ways on twelve thousand — none of them about prompt quality. Wiring a model into a real backend is a batch-processing problem with an expensive, rate-limited, non-deterministic function in the middle, and it needs everything batch processing always needed.
A conversation has a social latency budget, not a technical one — go past a second and people think the thing is broken. How a voice agent went from 1,530ms to 940ms to first audio, why every millisecond came from overlapping stages rather than speeding them up, the speculative optimisation that cost 38% more tokens for 60ms, and what a user hears when a dependency vanishes mid-sentence.
On a Tuesday my RAG system got quietly worse. Every request returned 200, no alert fired, and the answers stayed fluent and confident while being subtly about the wrong thing. Why AI systems fail silently, what a trace must actually record, why the mean is a lie, and the forty-minute root cause that would have taken days.
Three models, one laptop, thirty prompts, five runs each. Tokens per second, time to first token, memory and schema compliance for Llama 3.2 3B, Phi-4-mini and Mistral 7B — plus why I shipped the model that lost on speed, and what the Q4 to Q5 quantisation trade actually costs.
I opened a pull request that made my RAG system measurably worse and was completely certain it was an improvement. The build failed instead. What a golden evaluation pair actually contains, the four metrics that localise a failure, why 21% of my questions have no answer, how to pick a threshold from the noise floor, and the three of my own PRs it blocked.
My first agent made 41 tool calls to answer a question that needed two — and never threw a single error. The twelve lines an agent actually is, the five decisions inside the loop, why the tool schema is the real prompt, and the ten guardrails between a demo and something you would let near a customer.
A production RAG system with enforced citations, a local model benchmark, an observability layer, and a real-time voice agent — built in sequence as one system, measured end to end. Why I deliberately skipped fine-tuning, the numbers from every phase, and the four things I got wrong first.
What does it actually look like when a customer chats with a support bot and RAG plus an agentic loop run quietly behind the scenes? Four real-world automation blueprints — intelligent support, medical report analysis, legal consultation, and finance advisory — drawn as process diagrams you can build from.
The tooling around AI agents and RAG has exploded, and most "top 10 tools" lists are just logos. This is the opinionated map: what each tool is actually for, when to reach for it, when to skip it, and how the pieces fit into one coherent stack.
